Research Digest — 2026-11-07¶
Superionic Conductors & Solid Electrolytes¶
1. Temperature-Dependent Ion Dynamics in Argyrodite Superionic Conductor Li6PS5Cl with Fine-Tuned MACE Foundation Model¶
Source: arXiv:2611.05946 · 📅 2026-11-04 · ↗ Open paper
Molecular dynamics with a fine-tuned MACE foundation model, combined with essential dynamics analysis borrowed from biomolecular simulations, maps the temperature-dependent Li+ dynamics in argyrodite Li6PS5Cl. Below 500 K the Li trajectories are localized, phonon-like, and confined to a highly nonlinear low-dimensional manifold, while at ~800 K the dynamics become fully anharmonic; the 500-700 K window shows strong competition between the two regimes.
Relevance to DENG.Group
Directly relevant to Yanhao Deng's ML-potential work (fine-tuning a foundation model for a superionic conductor is exactly our workflow) and to our argyrodite/halide transport projects; the essential-dynamics trajectory analysis is a cheap new diagnostic we could apply to our own MD to separate vibrational from diffusive motion.
Amorphous Electrolytes¶
2. Collective dipole effects amplify field-driven ion transport in amorphous Li3PO4¶
Source: arXiv:2611.02617 · 📅 2026-11-02 · ↗ Open paper
MD with a machine-learned interatomic potential plus a machine-learned Born-effective-charge model measures Li mobility in amorphous Li3PO4 from the linear into the strongly nonlinear field regime. The amorphous phase shows far stronger field enhancement of transport than the crystal, traced to a collective dipole effect: the surrounding network moves with the hopping ion, making the field-work per hop broadly distributed, and since rates scale exponentially with that work the hops reorganize cooperatively.
Relevance to DENG.Group
Central to Tzu-Chen Liu's amorphous-electrolyte project: it gives a concrete microscopic mechanism (environment-mediated field-work distribution) for why amorphous electrolytes respond differently to fields, and demonstrates the ML Born-charge machinery we could reuse for our own amorphous systems. Also relevant to Naibing Wu's field-driven polymer transport modeling.
Polymer Electrolytes¶
3. PolyGPT: A Decoder-Only Transformer for Polymer Electrolyte Design¶
Source: arXiv:2611.07438 · 📅 2026-11-05 · ↗ Open paper
A 38.3M-parameter decoder-only Transformer (PolyGPT) generates polymer repeat units conditioned on property targets, requested chemical features, and an electrolyte-domain flag. Structural validity is excellent (99.5%), but electrolyte conditioning collapses output uniqueness from 99.99% to 51.7%, and adding the electrolyte condition only marginally improves the Tg prediction error (26.97 to 26.07 K) — a candid look at where generative design actually struggles.
Relevance to DENG.Group
Directly relevant to Naibing Wu's polymer-electrolyte simulation program: this is the inverse-design counterpart to our forward MD screening, and its honest error analysis (conditioning barely helps Tg prediction) is a useful calibration for how much to trust generative proposals before running our own simulations on them.
4. Apparent coordination-induced slowdown in PEO-based single-ion polymer electrolytes: A tracer-assisted Sticky Rouse model interpretation from molecular dynamics¶
Source: Polymer (10.1016/j.polymer.2026.130746) · 📅 2026-11-01 · ↗ Open paper
All-atom MD of PEO-based single-ion polymer electrolytes combined with tracer-particle analysis shows that apparent ion slowdown under strong coordination can be interpreted through a Sticky Rouse framework, connecting microscopic Li-solvent binding to chain-level relaxation dynamics.
Relevance to DENG.Group
Speaks to Naibing Wu's polymer-electrolyte work: the Sticky Rouse interpretation offers a coarse-grained bridge from MD trajectories to analytic transport models, which could help us rationalize ion-diffusion slowdown in our simulated single-ion conductors without rerunning full simulations at every composition.
ML Interatomic Potentials & Methods¶
5. Equivariant Deep Learning in Direct-Product Space for Electronic-Structure Hamiltonian Prediction¶
Source: arXiv:2611.07695 · 📅 2026-11-05 · ↗ Open paper
DP-Learner performs equivariant message passing directly on Hamiltonian-like operator blocks in the direct-product space of local orbital bases, rather than in the usual direct-sum space of irreducible spherical tensors. Using Lie-generator-based directional embeddings and invariantly weighted aggregation, it keeps Hamiltonian blocks contiguous, simplifies orbital-coupling manipulation, and improves hardware efficiency for electronic-structure prediction.
Relevance to DENG.Group
Methodological input for Yanhao Deng's ML-potential work and any future Hamiltonian/charge-screening models in the group: learning directly in the operator representation the target lives in is a design principle that could transfer to our equivariant force-field architectures, especially where long-range electrostatics matter.
Thermal Transport in Framework Materials¶
6. Heat Transport Mechanisms in MOF-5¶
Source: arXiv:2611.08024 · 📅 2026-11-05 · ↗ Open paper
Mode-projected MD spectra combined with spectral Wigner heat transport resolve two channels in MOF-5: low-frequency particle-like phonon propagation falls sharply from 200 to 500 K, while wave-like phonon coherence stays nearly constant and supplies about half the intrinsic conductivity at 500 K. Linker vibrations show split spectra signaling phonon-quasiparticle breakdown, a feature also found in ZIF-8 and HKUST-1, and an extrinsic scattering model reconciles calculated and measured conductivities.
Relevance to DENG.Group
Relevant to Jason Li's MOF adsorption/gas-kinetics work (heat management during adsorption) and to Mengke Li's thermal-property studies of framework-like halide electrolytes: the spectral Wigner methodology and the coherent/incoherent channel split are tools we could apply directly to our own framework and superionic systems.
Thin-Film Growth & Kinetics¶
7. Atomistic Insights into MOCVD Growth of GaAs using a Closed Loop Autonomous System Combining Directed Reinforcement Learning and Kinetic Monte Carlo Simulations¶
Source: arXiv:2611.01646 · 📅 2026-11-01 · ↗ Open paper
A closed-loop framework uses directed reinforcement learning to parameterize kinetic Monte Carlo simulations of MOCVD GaAs growth, circumventing the need for expensive DFT-derived rate constants while retaining atomistic resolution at experimental length and time scales. The RL agent steers KMC rate parameters toward agreement with target film morphology, making growth optimization autonomous rather than iterative and experimental.
Relevance to DENG.Group
Directly relevant to Jingyuan Lin's thin-film growth/kMC modeling: the RL-calibrated KMC loop is a concrete alternative to hand-derived rate catalogs for our growth simulations, and the workflow (cheap surrogate rates, closed-loop refinement) could accelerate our thin-film parameter searches.
Phase Field & Microstructure¶
8. Accelerating phase-field simulations of grain growth via tensor decomposition¶
Source: arXiv:2611.02118 · 📅 2026-11-01 · ↗ Open paper
A tensor-decomposition model-reduction framework exploits the low-rank structure of evolving phase-field variables to represent the high-dimensional solution in a spatially decomposed tensor format, cutting storage and the cost of repeatedly solving the phase-field equations for grain growth. The number of retained modes is controlled adaptively during evolution, enabling larger polycrystalline systems than conventional solvers.
Relevance to DENG.Group
Relevant to Cheng Peng's grain-boundary work in solid electrolytes and Shoutong Jin's phase-field dendrite simulations: the low-rank tensor solver is directly applicable to our large multi-variable phase-field problems, where memory cost currently limits system sizes and simulation times.